Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[gdansk]

Plus 5070 Ti is the cut down part on GB203, 5080 is the full part.

I'll also add that the full GB203 would be slower than the 5080 when at 304W instead of 360W. And possibly slower still if hobbled by equally slow memory, it'd be only 2/3rd the memory bandwidth to be comparable. I think it's close and I did not expect that.

 

[techjunkie123]

I'm not sure transistor counts are as comparable as people would like.
It's 2x the transistor density of Intel's B580 on what is supposed to be a similar density process. Is it really that many less transistors or are they counting differently?

N31 overall has a density of 110 MTr/mm^2. Comparing to just the N31 GCD on 5 nm, N48 has the same density (150 MTr/mm^2). But the GCD doesn't have infinity cache, memory controller, etc. This would suggest that the density improvements are quite dramatic, presumably for logic, given only minor improvements from the node.

 

[PJVol]

TBP, not TDP according to AMD site

Despite being technically different, they both refer to the same metric in the context of specs, i.e. max power allowed by the board's thermal design (read - cooler cap. + thermal limit)

 

[exquisitechar]

I'm not sure transistor counts are as comparable as people would like.
It's 2x the transistor density of Intel's B580 on what is supposed to be a similar density process. Is it really that many less transistors or are they counting differently?

Maybe they are, but either way, different designs can have massive differences in transistor counts despite having a similar die size using a similar node. “Performance per transistor” is a meaningless and irrelevant thing.

 

[Novacius]

Something I just now realized is that AMD is now ahead of NVIDIA in terms of performance/watt and performance/area for raster. Also possibly equal in RT. (NVIDIA has more cores so I suspect that is a big reason for the gap)

This is comparing the 9070XT to the 5070 Ti.

They were in the perfect position to drop a halo product. Of course the halo product got canned.

Huh? According to AMD the XT delivers 5070 Ti performance in raster at roughly the same power, meaning perf/W is now equal. Also the Ti is GB203's salvage, you have to compare to the 5080 for perf/area. GB203 is around 6% bigger than Navi48, but the 5080 should be around 15% faster, so perf/area of GB203 is slightly better.

N31 overall has a density of 110 MTr/mm^2. Comparing to just the N31 GCD on 5 nm, N48 has the same density (150 MTr/mm^2). But the GCD doesn't have infinity cache, memory controller, etc. This would suggest that the density improvements are quite dramatic, presumably for logic, given only minor improvements from the node.

According to TPU, Navi48 uses N4C, which should be denser than N4P.

 

[adroc_thurston]

so perf/area of GB203 is slightly better.

It has 40% moar membw.

5070 Ti performance in raster at roughly the same power, meaning perf/W is now equal

It haez 40% more membw too.

 

[exquisitechar]

Huh? According to AMD the XT delivers 5070 Ti performance in raster at roughly the same power, meaning perf/W is now equal. Also the Ti is GB203's salvage, you have to compare to the 5080 for perf/area. GB203 is around 6% bigger than Navi48, but the 5080 should be around 15% faster, so perf/area of GB203 is slightly better.

According to TPU, Navi48 uses N4C, which should be denser than N4P.

Although the 5080 is using GDDR7.

 

[Hail The Brain Slug]

No i mean its kingdom come 2 no RT in this game. Expected little higher.
But anyway judging by the tech power up game selection for it tests 5070ti will do decent.

From GameGPU: "The use of ray tracing technology makes the game even more cinematic, especially in scenes with sunsets and illuminated cities at night."

https://en.gamegpu.com/action-/-fps-/-tps/kingdom-come-delivery-ii-test-gpu-cpu

 

[Novacius]

It has 40% moar membw.

It haez 40% more membw too.

Somebody should bench the 5080 with downclocked memory then, to see how much that impacts performance

From GameGPU: "The use of ray tracing technology makes the game even more cinematic, especially in scenes with sunsets and illuminated cities at night."

https://en.gamegpu.com/action-/-fps-/-tps/kingdom-come-delivery-ii-test-gpu-cpu

That's software RT (SVOGI), there's no hardware acceleration happening

 

[adroc_thurston]

Somebody should bench the 5080 with downclocked memory then, to see how much that impacts performance

And with 300W TBP.

 

[gaav87]

Well got my hands on working non cheated driver time to break into some warehouse.

 

[gaav87]

China jd.com pre order prices converted:
No taichi or nitro+ prices enjoy.

 

[carrotmania]

China jd.com pre order prices converted:
No taichi or nitro+ prices enjoy.

View attachment 118516

You've got at least two $599 cards magically transformed into >850EUR. If you did this chart, can you give it a once, maybe twice, over?

 

[soresu]

I wonder if their improved h264 encoder is from Xilinx

At this point they've been a part of AMD so long that it doesn't make a difference.

 

[poke01]

Kingdom come 2 uses Voxel cone tracing, not RT. NV and Intel cards play well with this game.

Oh well at least AMD has CoD and Spider-Man.

 

[adroc_thurston]

NV and Intel cards play well with this game.

It's a cryengine thing. It hates AMD GPUs for idk why reasons.

 

[gaav87]

You've got at least two $599 cards magically transformed into >850EUR. If you did this chart, can you give it a once, maybe twice, over?

LoL

 

[SolidQ]

It's a cryengine thing. It hates AMD GPUs for idk why reasons.

ATI/AMD change many uarchs and still hates

 

[poke01]

It's a cryengine thing. It hates AMD GPUs for idk why reasons.

Yeah I know, just saying it’s doesn’t use hardware RT.

 

[poke01]

ATI/AMD change many uarchs and still hates

Happens, just like how NV will never catch up to AMD in CoD despite many different generations

 

[Mahboi]

At this point they've been a part of AMD so long that it doesn't make a difference.

No the h264 encoders in RDNA have always been butt, while Xilinx clearly had far better stuff for their Alveo transcoders. I was hoping that RDNA 4 would start implementing the Xilinx ones instead of the very poor AMD ones, I'm curious if that's the case.

 

[gdansk]

No the h264 encoders in RDNA have always been butt, while Xilinx clearly had far better stuff for their Alveo transcoders. I was hoping that RDNA 4 would start implementing the Xilinx ones instead of the very poor AMD ones, I'm curious if that's the case.

Most their open source software commits are from the same people as ever. I don't think there was a major change in teams, just new work.

 

[Mahboi]

Well, we'll see...I just am shocked to hear 20% differential after entire generations of tiny increments. 20% better takes their poor stuff up to NV level.

 

[Win2012R2]

RDNA is cool again, they'd be very stupid to change away from RDNA5

 

[IEC]

I'm still waiting for disappointment™.

But perhaps for the first time in a long time I might be disappointed to not be disappointed?